💡 We found that study designs supporting good inference use at least moderate cue set sizes and number of responses, as well as mixed or broad cue set types.

Read more in our open access paper:
doi.org/10.1371/jour...

⚠️ However, bias is high and varies a lot across study designs.

This renders comparisons of semantic network metrics between study designs very hard to interpret.

Semantic representation recovery was evaluated in terms of bias, resolution, and generalizability of semantic network metrics.

✅ Models of semantic representation can be inferred successfully from free association and relatedness judgments, if appropriate study designs are used!

How well do semantic networks infered from free association and relatedness judgments recover the semantic representations of simulated individuals?

We systematically varied four characteristics of empirical study designs: cue set type, cue set size, number of responses, and response type.

Check out the open access tutorial paper here: doi.org/10.5334/joc....

Check out the open source R package here: github.com/samuelae/ass...

To analyze group differences in the LLM-generated response frequencies from each cluster.

In the tutorial, we show, how we analyzed LLM-generated example free associations to the term 'intelligence', grouping responses into six clusters.

By prompting GPT-4 to produce responses typical for different genders, we compare it's gender-specific representation of 'intelligence'.

We demonstrate the associatoR workflow by analyzing GPT-4-generated free associations to the term 'intelligence'.